Harness

I keep a LEARNING.md retrospective in every repo I work on. Eighteen of them, 290 lessons, read exactly once. A deterministic script and one agent pass found seven failure shapes recurring across repos, and four of them became mechanical process changes the same day. Two of those shapes, I had already blogged about without recognizing them.

I stopped trusting my agent’s prose and started demanding evidence. Then a reviewer agent filed a confident, well-cited warning against code that no longer existed. Evidence has a grammar: freshness, correlation, independence. I watched all three rules break in one afternoon.